Cheatsheet: Many agents working together: multi-agent systems

The one idea

A multi-agent system is several specialized agents that coordinate. It helps some tasks and hurts others. The real question is fit, not ranking: do the jobs split cleanly enough that specialists beat one generalist, and is it worth the coordination cost?

What it is

Several agents, each with its own role, instructions, and focused tools, held together by either a coordinator that delegates or a pipeline of handoffs. Each agent is still just a model in a loop with tools; what is new is there are several, and they must communicate.

Advantages

• Specialization: a focused agent with a small, well-described toolbox picks reliably (the L4 tool-definition point at the system level).
• Parallelism: independent sub-tasks run at once.
• Modularity: build, test, and replace each agent independently.

The coordination cost (the omitted half)

• Communication overhead: context passed between agents can be lost or distorted at every handoff.
• More failure points: N agents plus the connections between them; only as reliable as the shakiest handoff.
• Who-decides-what: assigning work, resolving disagreement, knowing when done is itself hard.
• Debugging + latency: more places to trace, more rounds of communication.

Blunt consequence: many tasks are better as one well-designed generalist agent.

Specialists vs one generalist

Multiple agents win when	One generalist wins when
Task splits into distinct specialties	Task is mostly sequential
Sub-tasks parallelize	Pieces share too much context to hand off cleanly
One agent’s toolbox is so big it confuses itself	The task is small enough that coordination is pure overhead

One agent vs a team

TASK: short research brief.
GENERALIST: one loop, search -> read -> draft. One place to debug.
TEAM: coordinator -> researcher -> writer.
  Buys: tuned specialists, parallel across topics.
  Costs: 3 agents, 2 handoffs to garble findings, coordination logic, latency.

One brief: generalist wins. Fifty briefs a day across specialties: the team starts to pay.

Compound systems

Multi-agent is one shape of a compound system (models + retrievers + tools + agents composed together). Same rule for all: each component buys capability and costs coordination. Spend the tradeoff on purpose.

Pitfalls to dodge

• Reaching for multi-agent by default (start with one agent; split only when the task demands it).
• Ignoring coordination cost (specialization is visible; overhead is the hidden bill).
• Underestimating the coordination logic (who-does-what is a hard problem, not a free wrapper).
• Confusing more agents with more capability (more specialized + more complex, not smarter).
• Losing context at the seams (the typical failure is a bad handoff, not a bad agent).

Words to use precisely

• Multi-agent system: several coordinating agents, each specialized, working a larger task.
• Coordinator / orchestrator: an agent that splits work and delegates to specialists.
• Handoff: passing the task (and its context) from one agent to the next; the usual failure point.
• Compound system: an AI application built from multiple composed components, of which multi-agent is one shape.